ENH: missing value handling in basic statistics

[josef-pkt]

Currently we don't have much support for missing values outside of the model option for rowwise deletion.

#2630 is for improving descriptive statistics.

For univariate statistics like mean, std including robust statistics like mad or the quantile based skew and kurtosis, it would be better to have case/elementwise instead of rowwise deletion.

One problem is that simple vectorization will not work anymore if columns have nans/missing values in different rows. For simple statistics there are nan-aware functions, but we might need either a mask solution like MaskedArray statistics uses (essentially with 0/1 weights after changing to something finite) or looping over columns/series.

[rlucas7]

This CV post might also be relevant? (in tsa though)
https://stats.stackexchange.com/questions/381378/abbreviations-statsmodels-handling-nan